Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器

Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器

ID:18440501

大小:191.43 KB

页数:9页

时间:2018-09-18

Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器_第1页
Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器_第2页
Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器_第3页
Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器_第4页
Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器_第5页
资源描述:

《Ranking annotators for crowdsourced labeling tasks众源标记任务排序注释器》由会员上传分享,免费在线阅读,更多相关内容在学术论文-天天文库

1、RankingannotatorsforcrowdsourcedlabelingtasksVikasC.RaykarShipengYuSiemensHealthcare,Malvern,PA,USASiemensHealthcare,Malvern,PA,USAvikas.raykar@siemens.comshipeng.yu@siemens.comAbstractWiththeadventofcrowdsourcingservicesithasbecomequitecheapandreason-ablyeffectivetogetadatasetlab

2、eledbymultipleannotatorsinashortamountoftime.Variousmethodshavebeenproposedtoestimatetheconsensuslabelsbycorrectingforthebiasofannotatorswithdifferentkindsofexpertise.Oftenwehavelowqualityannotatorsorspammers–annotatorswhoassignlabelsrandomly(e.g.,withoutactuallylookingattheinstan

3、ce).Spammerscanmakethecostofacquiringlabelsveryexpensiveandcanpotentiallydegradethequalityofthecon-sensuslabels.Inthispaperweformalizethenotionofaspammeranddefineascorewhichcanbeusedtoranktheannotators—withthespammershavingascoreclosetozeroandthegoodannotatorshavingahighscorecloset

4、oone.1SpammersincrowdsourcedlabelingtasksAnnotatinganunlabeleddatasetisoneofthebottlenecksinusingsupervisedlearningtobuildgoodpredictivemodels.Gettingadatasetlabeledbyexpertscanbeexpensiveandtimeconsuming.Withtheadventofcrowdsourcingservices(Amazon'sMechanicalTurkbeingaprimeexampl

5、e)ithasbecomequiteeasyandinexpensivetoacquirelabelsfromalargenumberofannotatorsinashortamountoftime(see[8],[10],and[11]forsomecomputervisionandnaturallanguageprocessingcasestudies).Onedrawbackofmostcrowdsourcingservicesisthatwedonothavetightcontroloverthequalityoftheannotators.The

6、annotatorscancomefromadiversepoolincludinggenuineexperts,novices,biasedannotators,maliciousannotators,andspammers.Henceinordertogetgoodqualitylabelsrequestorstypicallygeteachinstancelabeledbymultipleannotatorsandthesemultipleannotationsarethenconsolidatedeitherusingasimplemajority

7、votingormoresophisticatedmeth-odsthatmodelandcorrectfortheannotatorbiases[3,9,6,7,14]and/ortaskcomplexity[2,13,12].Inthispaperweareinterestedinrankingannotatorsbasedonhowspammerlikeeachannotatoris.Inourcontextaspammerisalowqualityannotatorwhoassignsrandomlabels(maybebecausetheanno

8、tatordoesnotunderstandthelabeling

当前文档最多预览五页,下载文档查看全文

此文档下载收益归作者所有

当前文档最多预览五页,下载文档查看全文
温馨提示:
1. 部分包含数学公式或PPT动画的文件,查看预览时可能会显示错乱或异常,文件下载后无此问题,请放心下载。
2. 本文档由用户上传,版权归属用户,天天文库负责整理代发布。如果您对本文档版权有争议请及时联系客服。
3. 下载前请仔细阅读文档内容,确认文档内容符合您的需求后进行下载,若出现内容与标题不符可向本站投诉处理。
4. 下载文档时可能由于网络波动等原因无法下载或下载错误,付费完成后未能成功下载的用户请联系客服处理。